Choose one of three approaches. The steps for that option appear below.
Choose an option above. On a phone, swipe the tabs left or right to see all options →
Claude makes the evidence table as a spreadsheet and drafts the review. You correct between the stages.
You: Prepare input pack
Claude: Build evidence table
You: Correct row by row
Claude: Draft the review
You: Check and decide
Use: You
Have ready: the corrective action register, observation period, criteria, measurement definitions, follow-up coverage and attributable observations. Mark gaps as still to confirm.
Check nothing outside the supplied inputs is added.
Use: Claude
Ask Claude to build the action-to-measure evidence table as a spreadsheet: per action the supplied criteria, period, definition, comparator, coverage and missing-evidence flags.
What to paste into the tool
Task: Action-to-measure evidence table retaining supplied success criteria, observation period, measurement definitions, comparator where required, actual follow-up coverage and explicit missing/insufficient evidence. Use only the attached input pack. Keep every supplied criterion, period, definition and comparator exactly as given, and mark missing or insufficient evidence explicitly. Do not add anything beyond the supplied inputs.
Use: You
Check the table row by row against the original inputs and correct it in place.
Compare identifiers, definitions, dates and ambiguous claims against the original inputs.
Use: Claude
Ask Claude to draft the effectiveness review from the corrected table: implemented, observed, tested and unresolved actions, with unresolved fields and source references.
What to paste into the tool
Task: Draft effectiveness review separating implemented, observed, tested and unresolved actions, retaining supplied criteria, observation coverage and evidence limitations; Explicit unresolved fields and source references. Use only the corrected table and the supplied inputs. Separate implemented, observed, tested and unresolved actions, keep the supplied criteria and coverage limits visible, and list unresolved fields with their source references. Do not add facts beyond the pasted materials.
Use: You
Check the draft against the requirements; missing criteria or coverage stay marked as gaps. Confirm any next action separately.
Check the draft against every criterion, definition, comparator and actual coverage; task completion alone proves nothing.
You map each action to its criteria and coverage by hand. ChatGPT only drafts the review from that checked table.
You: Build table by hand
ChatGPT: Draft the review
You: Check against table
Use: You
Build the action-to-measure table by hand: per action, supplied criteria, observation period, measurement definition, comparator, actual coverage, source reference and evidence flag.
Compare identifiers, definitions, dates and ambiguous claims against the original inputs.
Use: ChatGPT
Paste the checked table and ask ChatGPT to draft the effectiveness review from it, keeping criteria and coverage limits visible.
What to paste into the tool
Task: Draft effectiveness review separating implemented, observed, tested and unresolved actions, retaining supplied criteria, observation coverage and evidence limitations; Explicit unresolved fields and source references. Use only the pasted checked table and the supplied inputs. Separate implemented, observed, tested and unresolved actions, keep the supplied criteria and coverage limits visible, and list unresolved fields with their source references. Do not add facts beyond the pasted materials.
Use: You
Compare the draft with the hand-built table line by line and fix any drift; confirm any next action separately. Missing criteria stay gaps.
Check the draft against every criterion, definition, comparator and actual coverage; task completion alone proves nothing.
Julius AI maps the observations to measures from your files. You check every row and write the review personally.
You: Select supplied inputs
Julius AI: Build evidence table
You: Check every row
You: Write the review
The second reviewer: Independent read
Use: You
Have ready: the register, observation period, criteria, measurement definitions, coverage and observations. Mark what is not known yet before uploading.
Check nothing outside the supplied inputs is added.
Use: Julius AI
Upload the prepared files and ask Julius AI to build the action-to-measure table with a flag where evidence is missing or insufficient.
What to paste into the tool
Task: Action-to-measure evidence table retaining supplied success criteria, observation period, measurement definitions, comparator where required, actual follow-up coverage and explicit missing/insufficient evidence. Upload the prepared files and build one table: per action, the supplied criteria, observation period, measurement definition, comparator where required, actual coverage, source reference, and a flag where evidence is missing or insufficient. Do not invent values.
Use: You
Check every row against the originals, correct it, and note what is still not known yet.
Compare identifiers, definitions, dates and ambiguous claims against the original inputs.
Use: You
Write the review yourself: implemented, observed, tested and unresolved actions. Mark actions with missing criteria or coverage as unresolved.
Check the draft against every criterion, definition, comparator and actual coverage; task completion alone proves nothing.
Use: The second reviewer
Read the draft independently against the originals and criteria; unsupported statements return to you as gaps. Confirm next actions separately.
Check the read was independent of your working notes. Return anything the originals do not support to the Review lead as a gap.
Best when: You want a downloadable evidence table, drafted review, one chat
What it does: Claude builds and drafts; you check twice
Strong at: The table comes back as a spreadsheet with formulas
Limit: A first-stage error can carry into the draft; check carefully
Free plan: Yes
Best when: You want to judge evidence adequacy personally
What it does: You build the table; ChatGPT drafts from it
Strong at: Every source ID and date is placed by you
Limit: First stage is hand work; slower on large inputs
Free plan: Yes, limited (uploads and analysis capped)
Best when: Observations sit in data files; you write the judgement
What it does: Julius AI tabulates; you check, write, get a second read
Strong at: You write the final draft; a second reviewer reads
Limit: Free files are deleted after an hour idle; download first
Free plan: Yes, limited (files stored 7 days only)
Before you paste into any AI tool, remove any confidential information. Put it back at the end.
Deciding between ChatGPT and Claude for this? See the comparison.
Named human compares every identifier, date, number and uncertain claim with the original inputs before the draft moves on.
Missing facts stay not stated; nothing is sent, published, assigned or authorised.
Cut down operational and personnel data pasted into any tool.
Check the result against your original inputs. Approve it before sharing or acting.